%
% (c) The GRASP/AQUA Project, Glasgow University, 1992-1998
%
\section[CoreRules]{Transformation rules}
\begin{code}
module Rules (
RuleBase,
emptyRuleBase, mkRuleBase, extendRuleBaseList,
unionRuleBase, pprRuleBase,
ruleCheckProgram,
mkSpecInfo, extendSpecInfo, addSpecInfo,
addIdSpecialisations,
rulesOfBinds, getRules, pprRulesForUser,
lookupRule, mkRule, roughTopNames
) where
#include "HsVersions.h"
import CoreSyn
import CoreSubst
import OccurAnal ( occurAnalyseExpr )
import CoreFVs ( exprFreeVars, exprsFreeVars, bindFreeVars, rulesFreeVars )
import CoreUtils ( exprType, eqExpr )
import PprCore ( pprRules )
import Type ( Type )
import TcType ( tcSplitTyConApp_maybe )
import Coercion
import CoreTidy ( tidyRules )
import Id
import IdInfo ( SpecInfo( SpecInfo ) )
import VarEnv
import VarSet
import Name ( Name, NamedThing(..) )
import NameEnv
import Unify ( ruleMatchTyX, MatchEnv(..) )
import BasicTypes ( Activation, CompilerPhase, isActive )
import StaticFlags ( opt_PprStyle_Debug )
import Outputable
import FastString
import Maybes
import Bag
import Util
import Data.List
\end{code}
Note [Overall plumbing for rules]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
* After the desugarer:
- The ModGuts initially contains mg_rules :: [CoreRule] of
locally-declared rules for imported Ids.
- Locally-declared rules for locally-declared Ids are attached to
the IdInfo for that Id. See Note [Attach rules to local ids] in
DsBinds
* TidyPgm strips off all the rules from local Ids and adds them to
mg_rules, so that the ModGuts has *all* the locally-declared rules.
* The HomePackageTable contains a ModDetails for each home package
module. Each contains md_rules :: [CoreRule] of rules declared in
that module. The HomePackageTable grows as ghc --make does its
up-sweep. In batch mode (ghc -c), the HPT is empty; all imported modules
are treated by the "external" route, discussed next, regardless of
which package they come from.
* The ExternalPackageState has a single eps_rule_base :: RuleBase for
Ids in other packages. This RuleBase simply grow monotonically, as
ghc --make compiles one module after another.
During simplification, interface files may get demand-loaded,
as the simplifier explores the unfoldings for Ids it has in
its hand. (Via an unsafePerformIO; the EPS is really a cache.)
That in turn may make the EPS rule-base grow. In contrast, the
HPT never grows in this way.
* The result of all this is that during Core-to-Core optimisation
there are four sources of rules:
(a) Rules in the IdInfo of the Id they are a rule for. These are
easy: fast to look up, and if you apply a substitution then
it'll be applied to the IdInfo as a matter of course.
(b) Rules declared in this module for imported Ids, kept in the
ModGuts. If you do a substitution, you'd better apply the
substitution to these. There are seldom many of these.
(c) Rules declared in the HomePackageTable. These never change.
(d) Rules in the ExternalPackageTable. These can grow in response
to lazy demand-loading of interfaces.
* At the moment (c) is carried in a reader-monad way by the CoreMonad.
The HomePackageTable doesn't have a single RuleBase because technically
we should only be able to "see" rules "below" this module; so we
generate a RuleBase for (c) by combing rules from all the modules
"below" us. That's why we can't just select the home-package RuleBase
from HscEnv.
[NB: we are inconsistent here. We should do the same for external
pacakges, but we don't. Same for type-class instances.]
* So in the outer simplifier loop, we combine (b-d) into a single
RuleBase, reading
(b) from the ModGuts,
(c) from the CoreMonad, and
(d) from its mutable variable
[Of coures this means that we won't see new EPS rules that come in
during a single simplifier iteration, but that probably does not
matter.]
%************************************************************************
%* *
\subsection[specialisation-IdInfo]{Specialisation info about an @Id@}
%* *
%************************************************************************
A @CoreRule@ holds details of one rule for an @Id@, which
includes its specialisations.
For example, if a rule for @f@ contains the mapping:
\begin{verbatim}
forall a b d. [Type (List a), Type b, Var d] ===> f' a b
\end{verbatim}
then when we find an application of f to matching types, we simply replace
it by the matching RHS:
\begin{verbatim}
f (List Int) Bool dict ===> f' Int Bool
\end{verbatim}
All the stuff about how many dictionaries to discard, and what types
to apply the specialised function to, are handled by the fact that the
Rule contains a template for the result of the specialisation.
There is one more exciting case, which is dealt with in exactly the same
way. If the specialised value is unboxed then it is lifted at its
definition site and unlifted at its uses. For example:
pi :: forall a. Num a => a
might have a specialisation
[Int#] ===> (case pi' of Lift pi# -> pi#)
where pi' :: Lift Int# is the specialised version of pi.
\begin{code}
mkRule :: Bool -> Bool -> RuleName -> Activation
-> Name -> [CoreBndr] -> [CoreExpr] -> CoreExpr -> CoreRule
mkRule is_auto is_local name act fn bndrs args rhs
= Rule { ru_name = name, ru_fn = fn, ru_act = act,
ru_bndrs = bndrs, ru_args = args,
ru_rhs = occurAnalyseExpr rhs,
ru_rough = roughTopNames args,
ru_auto = is_auto, ru_local = is_local }
roughTopNames :: [CoreExpr] -> [Maybe Name]
roughTopNames args = map roughTopName args
roughTopName :: CoreExpr -> Maybe Name
roughTopName (Type ty) = case tcSplitTyConApp_maybe ty of
Just (tc,_) -> Just (getName tc)
Nothing -> Nothing
roughTopName (Coercion _) = Nothing
roughTopName (App f _) = roughTopName f
roughTopName (Var f) | isGlobalId f
, isDataConWorkId f || idArity f > 0
= Just (idName f)
roughTopName _ = Nothing
ruleCantMatch :: [Maybe Name] -> [Maybe Name] -> Bool
ruleCantMatch (Just n1 : ts) (Just n2 : as) = n1 /= n2 || ruleCantMatch ts as
ruleCantMatch (_ : ts) (_ : as) = ruleCantMatch ts as
ruleCantMatch _ _ = False
\end{code}
Note [Care with roughTopName]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Consider this
module M where { x = a:b }
module N where { ...f x...
RULE f (p:q) = ... }
You'd expect the rule to match, because the matcher can
look through the unfolding of 'x'. So we must avoid roughTopName
returning 'M.x' for the call (f x), or else it'll say "can't match"
and we won't even try!!
However, suppose we have
RULE g (M.h x) = ...
foo = ...(g (M.k v))....
where k is a *function* exported by M. We never really match
functions (lambdas) except by name, so in this case it seems like
a good idea to treat 'M.k' as a roughTopName of the call.
\begin{code}
pprRulesForUser :: [CoreRule] -> SDoc
pprRulesForUser rules
= withPprStyle defaultUserStyle $
pprRules $
sortLe le_rule $
tidyRules emptyTidyEnv rules
where
le_rule r1 r2 = ru_name r1 <= ru_name r2
\end{code}
%************************************************************************
%* *
SpecInfo: the rules in an IdInfo
%* *
%************************************************************************
\begin{code}
mkSpecInfo :: [CoreRule] -> SpecInfo
mkSpecInfo rules = SpecInfo rules (rulesFreeVars rules)
extendSpecInfo :: SpecInfo -> [CoreRule] -> SpecInfo
extendSpecInfo (SpecInfo rs1 fvs1) rs2
= SpecInfo (rs2 ++ rs1) (rulesFreeVars rs2 `unionVarSet` fvs1)
addSpecInfo :: SpecInfo -> SpecInfo -> SpecInfo
addSpecInfo (SpecInfo rs1 fvs1) (SpecInfo rs2 fvs2)
= SpecInfo (rs1 ++ rs2) (fvs1 `unionVarSet` fvs2)
addIdSpecialisations :: Id -> [CoreRule] -> Id
addIdSpecialisations id []
= id
addIdSpecialisations id rules
= setIdSpecialisation id $
extendSpecInfo (idSpecialisation id) rules
rulesOfBinds :: [CoreBind] -> [CoreRule]
rulesOfBinds binds = concatMap (concatMap idCoreRules . bindersOf) binds
getRules :: RuleBase -> Id -> [CoreRule]
getRules rule_base fn
= idCoreRules fn ++ imp_rules
where
imp_rules = lookupNameEnv rule_base (idName fn) `orElse` []
\end{code}
Note [Where rules are found]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The rules for an Id come from two places:
(a) the ones it is born with, stored inside the Id iself (idCoreRules fn),
(b) rules added in other modules, stored in the global RuleBase (imp_rules)
It's tempting to think that
- LocalIds have only (a)
- non-LocalIds have only (b)
but that isn't quite right:
- PrimOps and ClassOps are born with a bunch of rules inside the Id,
even when they are imported
- The rules in PrelRules.builtinRules should be active even
in the module defining the Id (when it's a LocalId), but
the rules are kept in the global RuleBase
%************************************************************************
%* *
RuleBase
%* *
%************************************************************************
\begin{code}
type RuleBase = NameEnv [CoreRule]
emptyRuleBase :: RuleBase
emptyRuleBase = emptyNameEnv
mkRuleBase :: [CoreRule] -> RuleBase
mkRuleBase rules = extendRuleBaseList emptyRuleBase rules
extendRuleBaseList :: RuleBase -> [CoreRule] -> RuleBase
extendRuleBaseList rule_base new_guys
= foldl extendRuleBase rule_base new_guys
unionRuleBase :: RuleBase -> RuleBase -> RuleBase
unionRuleBase rb1 rb2 = plusNameEnv_C (++) rb1 rb2
extendRuleBase :: RuleBase -> CoreRule -> RuleBase
extendRuleBase rule_base rule
= extendNameEnv_Acc (:) singleton rule_base (ruleIdName rule) rule
pprRuleBase :: RuleBase -> SDoc
pprRuleBase rules = vcat [ pprRules (tidyRules emptyTidyEnv rs)
| rs <- nameEnvElts rules ]
\end{code}
%************************************************************************
%* *
Matching
%* *
%************************************************************************
\begin{code}
lookupRule :: (Activation -> Bool)
-> IdUnfoldingFun
-> InScopeSet
-> Id -> [CoreExpr]
-> [CoreRule] -> Maybe (CoreRule, CoreExpr)
lookupRule is_active id_unf in_scope fn args rules
=
case go [] rules of
[] -> Nothing
(m:ms) -> Just (findBest (fn,args) m ms)
where
rough_args = map roughTopName args
go :: [(CoreRule,CoreExpr)] -> [CoreRule] -> [(CoreRule,CoreExpr)]
go ms [] = ms
go ms (r:rs) = case (matchRule is_active id_unf in_scope args rough_args r) of
Just e -> go ((r,e):ms) rs
Nothing ->
go ms rs
findBest :: (Id, [CoreExpr])
-> (CoreRule,CoreExpr) -> [(CoreRule,CoreExpr)] -> (CoreRule,CoreExpr)
findBest _ (rule,ans) [] = (rule,ans)
findBest target (rule1,ans1) ((rule2,ans2):prs)
| rule1 `isMoreSpecific` rule2 = findBest target (rule1,ans1) prs
| rule2 `isMoreSpecific` rule1 = findBest target (rule2,ans2) prs
| debugIsOn = let pp_rule rule
| opt_PprStyle_Debug = ppr rule
| otherwise = doubleQuotes (ftext (ru_name rule))
in pprTrace "Rules.findBest: rule overlap (Rule 1 wins)"
(vcat [if opt_PprStyle_Debug then
ptext (sLit "Expression to match:") <+> ppr fn <+> sep (map ppr args)
else empty,
ptext (sLit "Rule 1:") <+> pp_rule rule1,
ptext (sLit "Rule 2:") <+> pp_rule rule2]) $
findBest target (rule1,ans1) prs
| otherwise = findBest target (rule1,ans1) prs
where
(fn,args) = target
isMoreSpecific :: CoreRule -> CoreRule -> Bool
isMoreSpecific (BuiltinRule {}) _ = False
isMoreSpecific (Rule {}) (BuiltinRule {}) = True
isMoreSpecific (Rule { ru_bndrs = bndrs1, ru_args = args1 })
(Rule { ru_bndrs = bndrs2, ru_args = args2 })
= isJust (matchN id_unfolding_fun in_scope bndrs2 args2 args1)
where
id_unfolding_fun _ = NoUnfolding
in_scope = mkInScopeSet (mkVarSet bndrs1)
noBlackList :: Activation -> Bool
noBlackList _ = False
\end{code}
Note [Extra args in rule matching]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
If we find a matching rule, we return (Just (rule, rhs)),
but the rule firing has only consumed as many of the input args
as the ruleArity says. It's up to the caller to keep track
of any left-over args. E.g. if you call
lookupRule ... f [e1, e2, e3]
and it returns Just (r, rhs), where r has ruleArity 2
then the real rewrite is
f e1 e2 e3 ==> rhs e3
You might think it'd be cleaner for lookupRule to deal with the
leftover arguments, by applying 'rhs' to them, but the main call
in the Simplifier works better as it is. Reason: the 'args' passed
to lookupRule are the result of a lazy substitution
\begin{code}
matchRule :: (Activation -> Bool) -> IdUnfoldingFun
-> InScopeSet
-> [CoreExpr] -> [Maybe Name]
-> CoreRule -> Maybe CoreExpr
matchRule _is_active id_unf _in_scope args _rough_args
(BuiltinRule { ru_try = match_fn })
= case match_fn id_unf args of
Just expr -> Just expr
Nothing -> Nothing
matchRule is_active id_unf in_scope args rough_args
(Rule { ru_act = act, ru_rough = tpl_tops,
ru_bndrs = tpl_vars, ru_args = tpl_args,
ru_rhs = rhs })
| not (is_active act) = Nothing
| ruleCantMatch tpl_tops rough_args = Nothing
| otherwise
= case matchN id_unf in_scope tpl_vars tpl_args args of
Nothing -> Nothing
Just (bind_wrapper, tpl_vals) -> Just (bind_wrapper $
rule_fn `mkApps` tpl_vals)
where
rule_fn = occurAnalyseExpr (mkLams tpl_vars rhs)
matchN :: IdUnfoldingFun
-> InScopeSet
-> [Var]
-> [CoreExpr]
-> [CoreExpr]
-> Maybe (BindWrapper,
[CoreExpr])
matchN id_unf in_scope tmpl_vars tmpl_es target_es
= do { subst <- go init_menv emptyRuleSubst tmpl_es target_es
; return (rs_binds subst,
map (lookup_tmpl subst) tmpl_vars') }
where
(init_rn_env, tmpl_vars') = mapAccumL rnBndrL (mkRnEnv2 in_scope) tmpl_vars
init_menv = RV { rv_tmpls = mkVarSet tmpl_vars', rv_lcl = init_rn_env
, rv_fltR = mkEmptySubst (rnInScopeSet init_rn_env)
, rv_unf = id_unf }
go _ subst [] _ = Just subst
go _ _ _ [] = Nothing
go menv subst (t:ts) (e:es) = do { subst1 <- match menv subst t e
; go menv subst1 ts es }
lookup_tmpl :: RuleSubst -> Var -> CoreExpr
lookup_tmpl (RS { rs_tv_subst = tv_subst, rs_id_subst = id_subst }) tmpl_var'
| isId tmpl_var' = case lookupVarEnv id_subst tmpl_var' of
Just e -> e
_ -> unbound tmpl_var'
| otherwise = case lookupVarEnv tv_subst tmpl_var' of
Just ty -> Type ty
Nothing -> unbound tmpl_var'
unbound var = pprPanic "Template variable unbound in rewrite rule"
(ppr var $$ ppr tmpl_vars $$ ppr tmpl_vars' $$ ppr tmpl_es $$ ppr target_es)
\end{code}
Note [Template binders]
~~~~~~~~~~~~~~~~~~~~~~~
Consider the following match:
Template: forall x. f x
Target: f (x+1)
This should succeed, because the template variable 'x' has
nothing to do with the 'x' in the target.
On reflection, this case probably does just work, but this might not
Template: forall x. f (\x.x)
Target: f (\y.y)
Here we want to clone when we find the \x, but to know that x must be in scope
To achive this, we use rnBndrL to rename the template variables if
necessary; the renamed ones are the tmpl_vars'
%************************************************************************
%* *
The main matcher
%* *
%************************************************************************
---------------------------------------------
The inner workings of matching
---------------------------------------------
\begin{code}
data RuleEnv = RV { rv_tmpls :: VarSet
, rv_lcl :: RnEnv2
, rv_fltR :: Subst
, rv_unf :: IdUnfoldingFun
}
data RuleSubst = RS { rs_tv_subst :: TvSubstEnv
, rs_id_subst :: IdSubstEnv
, rs_binds :: BindWrapper
, rs_bndrs :: VarSet
}
type BindWrapper = CoreExpr -> CoreExpr
emptyRuleSubst :: RuleSubst
emptyRuleSubst = RS { rs_tv_subst = emptyVarEnv, rs_id_subst = emptyVarEnv
, rs_binds = \e -> e, rs_bndrs = emptyVarSet }
match :: RuleEnv
-> RuleSubst
-> CoreExpr
-> CoreExpr
-> Maybe RuleSubst
match renv subst (Var v1) e2 = match_var renv subst v1 e2
match renv subst (Note _ e1) e2 = match renv subst e1 e2
match renv subst e1 (Note _ e2) = match renv subst e1 e2
match renv subst e1 (Var v2)
| not (inRnEnvR rn_env v2)
, Just e2' <- expandUnfolding_maybe (rv_unf renv v2')
= match (renv { rv_lcl = nukeRnEnvR rn_env }) subst e1 e2'
where
v2' = lookupRnInScope rn_env v2
rn_env = rv_lcl renv
match renv subst e1 (Let bind e2)
| okToFloat (rv_lcl renv) (bindFreeVars bind)
= match (renv { rv_fltR = flt_subst' })
(subst { rs_binds = rs_binds subst . Let bind'
, rs_bndrs = extendVarSetList (rs_bndrs subst) new_bndrs })
e1 e2
where
flt_subst = addInScopeSet (rv_fltR renv) (rs_bndrs subst)
(flt_subst', bind') = substBind flt_subst bind
new_bndrs = bindersOf bind'
match _ subst (Lit lit1) (Lit lit2)
| lit1 == lit2
= Just subst
match renv subst (App f1 a1) (App f2 a2)
= do { subst' <- match renv subst f1 f2
; match renv subst' a1 a2 }
match renv subst (Lam x1 e1) (Lam x2 e2)
= match renv' subst e1 e2
where
renv' = renv { rv_lcl = rnBndr2 (rv_lcl renv) x1 x2
, rv_fltR = delBndr (rv_fltR renv) x2 }
match renv subst (Lam x1 e1) e2
= match renv' subst e1 (App e2 (varToCoreExpr new_x))
where
(rn_env', new_x) = rnEtaL (rv_lcl renv) x1
renv' = renv { rv_lcl = rn_env' }
match renv subst e1 (Lam x2 e2)
= match renv' subst (App e1 (varToCoreExpr new_x)) e2
where
(rn_env', new_x) = rnEtaR (rv_lcl renv) x2
renv' = renv { rv_lcl = rn_env' }
match renv subst (Case e1 x1 ty1 alts1) (Case e2 x2 ty2 alts2)
= do { subst1 <- match_ty renv subst ty1 ty2
; subst2 <- match renv subst1 e1 e2
; let renv' = rnMatchBndr2 renv subst x1 x2
; match_alts renv' subst2 alts1 alts2
}
match renv subst (Type ty1) (Type ty2)
= match_ty renv subst ty1 ty2
match renv subst (Coercion co1) (Coercion co2)
= match_co renv subst co1 co2
match renv subst (Cast e1 co1) (Cast e2 co2)
= do { subst1 <- match_co renv subst co1 co2
; match renv subst1 e1 e2 }
match _ _ _e1 _e2 =
Nothing
match_co :: RuleEnv
-> RuleSubst
-> Coercion
-> Coercion
-> Maybe RuleSubst
match_co renv subst (CoVarCo cv) co
= match_var renv subst cv (Coercion co)
match_co _ _ co1 _
= pprTrace "match_co baling out" (ppr co1) Nothing
rnMatchBndr2 :: RuleEnv -> RuleSubst -> Var -> Var -> RuleEnv
rnMatchBndr2 renv subst x1 x2
= renv { rv_lcl = rnBndr2 rn_env x1 x2
, rv_fltR = delBndr (rv_fltR renv) x2 }
where
rn_env = addRnInScopeSet (rv_lcl renv) (rs_bndrs subst)
match_alts :: RuleEnv
-> RuleSubst
-> [CoreAlt]
-> [CoreAlt]
-> Maybe RuleSubst
match_alts _ subst [] []
= return subst
match_alts renv subst ((c1,vs1,r1):alts1) ((c2,vs2,r2):alts2)
| c1 == c2
= do { subst1 <- match renv' subst r1 r2
; match_alts renv subst1 alts1 alts2 }
where
renv' = foldl mb renv (vs1 `zip` vs2)
mb renv (v1,v2) = rnMatchBndr2 renv subst v1 v2
match_alts _ _ _ _
= Nothing
okToFloat :: RnEnv2 -> VarSet -> Bool
okToFloat rn_env bind_fvs
= foldVarSet ((&&) . not_captured) True bind_fvs
where
not_captured fv = not (inRnEnvR rn_env fv)
match_var :: RuleEnv
-> RuleSubst
-> Var
-> CoreExpr
-> Maybe RuleSubst
match_var renv@(RV { rv_tmpls = tmpls, rv_lcl = rn_env, rv_fltR = flt_env })
subst v1 e2
| v1' `elemVarSet` tmpls
= match_tmpl_var renv subst v1' e2
| otherwise
= case e2 of
Var v2 | v1' == rnOccR rn_env v2
-> Just subst
| Var v2' <- lookupIdSubst (text "match_var") flt_env v2
, v1' == v2'
-> Just subst
_ -> Nothing
where
v1' = rnOccL rn_env v1
match_tmpl_var :: RuleEnv
-> RuleSubst
-> Var
-> CoreExpr
-> Maybe RuleSubst
match_tmpl_var renv@(RV { rv_lcl = rn_env, rv_fltR = flt_env })
subst@(RS { rs_id_subst = id_subst, rs_bndrs = let_bndrs })
v1' e2
| any (inRnEnvR rn_env) (varSetElems (exprFreeVars e2))
= Nothing
| Just e1' <- lookupVarEnv id_subst v1'
= if eqExpr (rnInScopeSet rn_env) e1' e2'
then Just subst
else Nothing
| otherwise
=
do { subst' <- match_ty renv subst (idType v1') (exprType e2)
; return (subst' { rs_id_subst = id_subst' }) }
where
e2' | isEmptyVarSet let_bndrs = e2
| otherwise = substExpr (text "match_tmpl_var") flt_env e2
id_subst' = extendVarEnv (rs_id_subst subst) v1' e2'
match_ty :: RuleEnv
-> RuleSubst
-> Type
-> Type
-> Maybe RuleSubst
match_ty renv subst ty1 ty2
= do { tv_subst' <- Unify.ruleMatchTyX menv tv_subst ty1 ty2
; return (subst { rs_tv_subst = tv_subst' }) }
where
tv_subst = rs_tv_subst subst
menv = ME { me_tmpls = rv_tmpls renv, me_env = rv_lcl renv }
\end{code}
Note [Expanding variables]
~~~~~~~~~~~~~~~~~~~~~~~~~~
Here is another Very Important rule: if the term being matched is a
variable, we expand it so long as its unfolding is "expandable". (Its
occurrence information is not necessarily up to date, so we don't use
it.) By "expandable" we mean a WHNF or a "constructor-like" application.
This is the key reason for "constructor-like" Ids. If we have
{-# NOINLINE [1] CONLIKE g #-}
{-# RULE f (g x) = h x #-}
then in the term
let v = g 3 in ....(f v)....
we want to make the rule fire, to replace (f v) with (h 3).
Note [Do not expand locally-bound variables]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Do *not* expand locally-bound variables, else there's a worry that the
unfolding might mention variables that are themselves renamed.
Example
case x of y { (p,q) -> ...y... }
Don't expand 'y' to (p,q) because p,q might themselves have been
renamed. Essentially we only expand unfoldings that are "outside"
the entire match.
Hence, (a) the guard (not (isLocallyBoundR v2))
(b) when we expand we nuke the renaming envt (nukeRnEnvR).
Note [Notes in RULE matching]
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Look through Notes in both template and expression being matched. In
particular, we don't want to be confused by InlineMe notes. Maybe we
should be more careful about profiling notes, but for now I'm just
riding roughshod over them. cf Note [Notes in call patterns] in
SpecConstr
Note [Matching lets]
~~~~~~~~~~~~~~~~~~~~
Matching a let-expression. Consider
RULE forall x. f (g x) =
and target expression
f (let { w=R } in g E))
Then we'd like the rule to match, to generate
let { w=R } in (\x. ) E
In effect, we want to float the let-binding outward, to enable
the match to happen. This is the WHOLE REASON for accumulating
bindings in the RuleSubst
We can only do this if the free variables of R are not bound by the
part of the target expression outside the let binding; e.g.
f (\v. let w = v+1 in g E)
Here we obviously cannot float the let-binding for w. Hence the
use of okToFloat.
There are a couple of tricky points.
(a) What if floating the binding captures a variable?
f (let v = x+1 in v) v
--> NOT!
let v = x+1 in f (x+1) v
(b) What if two non-nested let bindings bind the same variable?
f (let v = e1 in b1) (let v = e2 in b2)
--> NOT!
let v = e1 in let v = e2 in (f b2 b2)
See testsuite test "RuleFloatLet".
Our cunning plan is this:
* Along with the growing substitution for template variables
we maintain a growing set of floated let-bindings (rs_binds)
plus the set of variables thus bound.
* The RnEnv2 in the MatchEnv binds only the local binders
in the term (lambdas, case)
* When we encounter a let in the term to be matched, we
check that does not mention any locally bound (lambda, case)
variables. If so we fail
* We use CoreSubst.substBind to freshen the binding, using an
in-scope set that is the original in-scope variables plus the
rs_bndrs (currently floated let-bindings). So in (a) above
we'll freshen the 'v' binding; in (b) above we'll freshen
the *second* 'v' binding.
* We apply that freshening substitution, in a lexically-scoped
way to the term, although lazily; this is the rv_fltR field.
Note [Matching cases]
~~~~~~~~~~~~~~~~~~~~~
{- NOTE: This idea is currently disabled. It really only works if
the primops involved are OkForSpeculation, and, since
they have side effects readIntOfAddr and touch are not.
Maybe we'll get back to this later . -}
Consider
f (case readIntOffAddr# p# i# realWorld# of { (# s#, n# #) ->
case touch# fp s# of { _ ->
I# n# } } )
This happened in a tight loop generated by stream fusion that
Roman encountered. We'd like to treat this just like the let
case, because the primops concerned are ok-for-speculation.
That is, we'd like to behave as if it had been
case readIntOffAddr# p# i# realWorld# of { (# s#, n# #) ->
case touch# fp s# of { _ ->
f (I# n# } } )
Note [Lookup in-scope]
~~~~~~~~~~~~~~~~~~~~~~
Consider this example
foo :: Int -> Maybe Int -> Int
foo 0 (Just n) = n
foo m (Just n) = foo (m-n) (Just n)
SpecConstr sees this fragment:
case w_smT of wild_Xf [Just A] {
Data.Maybe.Nothing -> lvl_smf;
Data.Maybe.Just n_acT [Just S(L)] ->
case n_acT of wild1_ams [Just A] { GHC.Base.I# y_amr [Just L] ->
\$wfoo_smW (GHC.Prim.-# ds_Xmb y_amr) wild_Xf
}};
and correctly generates the rule
RULES: "SC:$wfoo1" [0] __forall {y_amr [Just L] :: GHC.Prim.Int#
sc_snn :: GHC.Prim.Int#}
\$wfoo_smW sc_snn (Data.Maybe.Just @ GHC.Base.Int (GHC.Base.I# y_amr))
= \$s\$wfoo_sno y_amr sc_snn ;]
BUT we must ensure that this rule matches in the original function!
Note that the call to \$wfoo is
\$wfoo_smW (GHC.Prim.-# ds_Xmb y_amr) wild_Xf
During matching we expand wild_Xf to (Just n_acT). But then we must also
expand n_acT to (I# y_amr). And we can only do that if we look up n_acT
in the in-scope set, because in wild_Xf's unfolding it won't have an unfolding
at all.
That is why the 'lookupRnInScope' call in the (Var v2) case of 'match'
is so important.
%************************************************************************
%* *
Rule-check the program
%* *
%************************************************************************
We want to know what sites have rules that could have fired but didn't.
This pass runs over the tree (without changing it) and reports such.
\begin{code}
ruleCheckProgram :: CompilerPhase
-> String
-> RuleBase
-> [CoreBind]
-> SDoc
ruleCheckProgram phase rule_pat rule_base binds
| isEmptyBag results
= text "Rule check results: no rule application sites"
| otherwise
= vcat [text "Rule check results:",
line,
vcat [ p $$ line | p <- bagToList results ]
]
where
env = RuleCheckEnv { rc_is_active = isActive phase
, rc_id_unf = idUnfolding
, rc_pattern = rule_pat
, rc_rule_base = rule_base }
results = unionManyBags (map (ruleCheckBind env) binds)
line = text (replicate 20 '-')
data RuleCheckEnv = RuleCheckEnv {
rc_is_active :: Activation -> Bool,
rc_id_unf :: IdUnfoldingFun,
rc_pattern :: String,
rc_rule_base :: RuleBase
}
ruleCheckBind :: RuleCheckEnv -> CoreBind -> Bag SDoc
ruleCheckBind env (NonRec _ r) = ruleCheck env r
ruleCheckBind env (Rec prs) = unionManyBags [ruleCheck env r | (_,r) <- prs]
ruleCheck :: RuleCheckEnv -> CoreExpr -> Bag SDoc
ruleCheck _ (Var _) = emptyBag
ruleCheck _ (Lit _) = emptyBag
ruleCheck _ (Type _) = emptyBag
ruleCheck _ (Coercion _) = emptyBag
ruleCheck env (App f a) = ruleCheckApp env (App f a) []
ruleCheck env (Note _ e) = ruleCheck env e
ruleCheck env (Cast e _) = ruleCheck env e
ruleCheck env (Let bd e) = ruleCheckBind env bd `unionBags` ruleCheck env e
ruleCheck env (Lam _ e) = ruleCheck env e
ruleCheck env (Case e _ _ as) = ruleCheck env e `unionBags`
unionManyBags [ruleCheck env r | (_,_,r) <- as]
ruleCheckApp :: RuleCheckEnv -> Expr CoreBndr -> [Arg CoreBndr] -> Bag SDoc
ruleCheckApp env (App f a) as = ruleCheck env a `unionBags` ruleCheckApp env f (a:as)
ruleCheckApp env (Var f) as = ruleCheckFun env f as
ruleCheckApp env other _ = ruleCheck env other
\end{code}
\begin{code}
ruleCheckFun :: RuleCheckEnv -> Id -> [CoreExpr] -> Bag SDoc
ruleCheckFun env fn args
| null name_match_rules = emptyBag
| otherwise = unitBag (ruleAppCheck_help env fn args name_match_rules)
where
name_match_rules = filter match (getRules (rc_rule_base env) fn)
match rule = (rc_pattern env) `isPrefixOf` unpackFS (ruleName rule)
ruleAppCheck_help :: RuleCheckEnv -> Id -> [CoreExpr] -> [CoreRule] -> SDoc
ruleAppCheck_help env fn args rules
=
vcat [text "Expression:" <+> ppr (mkApps (Var fn) args),
vcat (map check_rule rules)]
where
n_args = length args
i_args = args `zip` [1::Int ..]
rough_args = map roughTopName args
check_rule rule = rule_herald rule <> colon <+> rule_info rule
rule_herald (BuiltinRule { ru_name = name })
= ptext (sLit "Builtin rule") <+> doubleQuotes (ftext name)
rule_herald (Rule { ru_name = name })
= ptext (sLit "Rule") <+> doubleQuotes (ftext name)
rule_info rule
| Just _ <- matchRule noBlackList (rc_id_unf env) emptyInScopeSet args rough_args rule
= text "matches (which is very peculiar!)"
rule_info (BuiltinRule {}) = text "does not match"
rule_info (Rule { ru_act = act,
ru_bndrs = rule_bndrs, ru_args = rule_args})
| not (rc_is_active env act) = text "active only in later phase"
| n_args < n_rule_args = text "too few arguments"
| n_mismatches == n_rule_args = text "no arguments match"
| n_mismatches == 0 = text "all arguments match (considered individually), but rule as a whole does not"
| otherwise = text "arguments" <+> ppr mismatches <+> text "do not match (1-indexing)"
where
n_rule_args = length rule_args
n_mismatches = length mismatches
mismatches = [i | (rule_arg, (arg,i)) <- rule_args `zip` i_args,
not (isJust (match_fn rule_arg arg))]
lhs_fvs = exprsFreeVars rule_args
match_fn rule_arg arg = match renv emptyRuleSubst rule_arg arg
where
in_scope = mkInScopeSet (lhs_fvs `unionVarSet` exprFreeVars arg)
renv = RV { rv_lcl = mkRnEnv2 in_scope
, rv_tmpls = mkVarSet rule_bndrs
, rv_fltR = mkEmptySubst in_scope
, rv_unf = rc_id_unf env }
\end{code}